feat(v7-h07): per_brick_grad_norms in train extras + 1/1 pytest

Author: apstenku123

cppmega_v4/runner/stages.py

@@ -1747,6 +1747,8 @@ def _base(k: str) -> str | None:
                     "top1_token_drift": top1_token_drift,
                 },
                 "side_channels_observed": side_channels_observed,
+                "per_brick_grad_norms": _safe_per_brick_grads(
+                    all_modules, grads if 'grads' in dir() else None),
                 "side_channels_forward_effect": {
                     "doc_ids_mask_density": sc_doc_ids_mask_density,
                     "doc_mask_applied": sc_doc_mask_applied,
@@ -2174,6 +2176,17 @@ def read_ckpt_metadata(path: str) -> dict | None:
     return out or None
 
 
+def _safe_per_brick_grads(model: Any, grads: Any) -> dict[str, float]:
+    """V7-H07: extras.per_brick_grad_norms — best-effort wrapper."""
+    if grads is None:
+        return {}
+    try:
+        from cppmega_v4.runtime.per_brick_probes import per_brick_grad_norms
+        return per_brick_grad_norms(model, grads)
+    except Exception:
+        return {}
+
+
 def _ema_smooth(values: list[float], window: int = 10) -> list[float]:
     """G15: simple windowed-mean smoothing (cheaper than true EMA, same
     job for visualising convergence)."""

tests/v4/test_per_brick_grads_extras.py

@@ -0,0 +1,43 @@
+"""V7-H07: extras.per_brick_grad_norms surfaces real per-layer grads."""
+
+from __future__ import annotations
+
+from cppmega_v4.jsonrpc.schema import VerifyParams
+from cppmega_v4.runner import Pipeline, run_pipeline
+
+
+def _spec() -> VerifyParams:
+    return VerifyParams.model_validate({
+        "graph": {
+            "nodes": [
+                {"id": "attn", "kind": "attention",
+                 "params": {"num_heads": 4, "head_dim": 64}},
+                {"id": "mlp", "kind": "mlp", "params": {}},
+            ],
+            "edges": [{"src": "attn", "dst": "mlp"}],
+        },
+        "dim_env": {"B": 1, "S": 8, "H": 128,
+                    "nh": 2, "nkv": 1, "head_dim": 64},
+        "loss": {"kind": "cross_entropy", "head_outputs": ["mlp"]},
+        "optim": {"kind": "adamw",
+                  "groups": [{"matcher": "all", "lr": 1e-3,
+                              "weight_decay": 0.01,
+                              "betas": [0.9, 0.95]}]},
+    })
+
+
+def test_v7_h07_extras_carries_per_brick_grad_norms():
+    rep = run_pipeline(_spec(), Pipeline.from_dict({
+        "stages": ["parse", "verify_build_spec", "build_model", "train"],
+        "stage_options": {"train": {"num_steps": 2}},
+    }))
+    tr = next(s for s in rep.stages if s.name == "train")
+    pbg = tr.extras["per_brick_grad_norms"]
+    assert isinstance(pbg, dict)
+    assert len(pbg) >= 1
+    for k, v in pbg.items():
+        assert k.startswith("layers.") or k in (
+            "shared_expert", "train_token_embedding",
+            "side_channel_token_embedding",
+        )
+        assert v > 0